In order to improve the prediction accuracy of pedestrian crossing patterns by conventional vehicles in unsignalized crosswalk road sections, a pedestrian crossing pattern prediction model integrating extreme gradient boosting (XGBoost) and multilayer perceptron (MLP) algorithms was proposed. First, the pedestrian-vehicle interaction data in the unsignalized crosswalk section were collected based on the cameras and LiDAR installed on the roadside, and the behavioral characteristics of pedestrians and vehicles were analyzed, and then the factors affecting the pedestrian crossing patterns were screened. Next, the predictive effects of different combinations when used as model inputs were explored. Finally, vehicle speed, vehicle-to-zebra crossing distance, time to collision(TTC) and pedestrian step speed were used as model inputs, and pedestrian crossing patterns were categorized into direct crossing and waiting crossing and used as model outputs, and the XGBoost-MLP model for pedestrian crossing pattern prediction was established. The prediction accuracy of this model for pedestrian crossing patterns reaches 88.65%, which compares with the single XGBoost model and the MLP model, and its accuracy is improved by 3.85% and 2.61% compared to the single XGBoost model and MLP model, respectively.